Elevation Map

A 2.5D terrain map that divides the ground into a grid and stores one height value per cell, commonly used by legged robots.

An elevation map is a 2.5D map: the ground around a robot is divided into a regular grid on the horizontal plane, and each cell stores just one height value (usually with a variance representing uncertainty too), rather than a full 3D voxel grid. The data comes from range sensors like depth cameras and lidar, continuously fused and updated together with the robot’s pose estimate. elevation_mapping, developed starting in 2014 by Péter Fankhauser and colleagues at ETH Zurich, is a commonly used open-source implementation that builds the map robot-centered and explicitly accounts for pose-drift uncertainty; it is no longer maintained. The 2022 elevation_mapping_cupy moved the computation to the GPU and added traversability, semantic, and other layers. An elevation map is more compact and faster to query than a point cloud, and perceptive locomotion for quadruped and humanoid robots commonly reads terrain height in a patch around the feet from it as policy input. Its limitation is that each cell holds only one height, so it can’t represent overhangs like the underside of a table or a bridge.

ExampleWhen training perceptive legged locomotion in Isaac Lab, a grid of terrain heights around the robot (a height scan) is commonly used as an observation; once deployed on the real robot, these heights are looked up from an elevation map built in real time.

Also called

2.5D Elevation Map, Elevation Mapping, Height Map

Related

Height Scan · Traversability Estimation · Perceptive Locomotion · Occupancy Grid Map · Rough-terrain Locomotion · ANYbotics ANYmal

Sources

ANYbotics/elevation_mapping (GitHub)
leggedrobotics/elevation_mapping_cupy (GitHub)